Intelligent storyboard script generation system and method

Abstract

The present invention discloses an intelligent storyboard script generation system and method, which belongs to the technical field of software engineering, including a front-end interactive interface layer, a communication layer one, a back-end business logic layer, a communication layer two and a computing layer connected in sequence, wherein the front-end interactive interface layer includes text editing and image preview functions, the back-end business logic layer includes a text generation module composed of a large language model (LLM) and an image generation module composed of a text-to-image generation class model (Stable-diffusion), the text generation module and the image generation module process input content to obtain a computationally intensive task, and the computing layer is set as a cluster composed of several servers to calculate the computationally intensive task. The present invention constructs a storyboard script generation system with coherent narrative capabilities and support for open source technology, a text generation function that meets the user's storyboard text creation needs, and can generate images.

Description

Technical Field

[0001] The present invention relates to the technical field of software engineering, and in particular to an intelligent storyboard script generation system and method. Background Art

[0002] Large Language Models (LLMs) are an important research direction in the field of artificial intelligence in recent years. The core of LLMs is the Transformer model, which processes sequence data through a self-attention mechanism and adopts an autoregressive pre-training strategy to consider the information of all previous words when predicting the next word. Vaswani et al. proposed the Transformer in 2017, which used the self-attention mechanism to model long-distance dependencies, thereby effectively capturing the contextual information in the text sequence and improving the performance of the language model. The introduction of this model laid the foundation for the subsequent development of large language models.

[0003] In 2018, Google released BERT, a language model with a bidirectional encoder-decoder structure. It uses the Transformer model to perform pre-training on large-scale text data, and achieves the encoding of bidirectional context, thereby achieving significant performance improvements in multiple natural language processing tasks. The success of BERT marks the gradual popularization and application of large language models in the field of natural language processing. Subsequently, OpenAI released the GPT series of models, which adopted a unidirectional autoregressive pre-training strategy, performed pre-training on large-scale text data, and achieved the generation of future text sequences.

[0004] Although large language models have achieved excellent performance on a variety of tasks, they are still constrained by inherent limitations, such as: (1) Transience: LLMs inherently lack persistent memory or state, requiring additional software or systems to maintain and manage context. (2) Probabilistic nature: The stochastic nature of LLMs leads to variable responses. (3) Staleness: The reliance on pre-training data limits LLMs to using only historical knowledge and cannot perceive or update in real time. (4) Hallucination: LLMs may generate information that appears reasonable but is in fact incorrect. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent storyboard script generation system and method. In view of the inherent limitations of large language models such as transient, probabilistic, outdated and hallucinatory, a storyboard script generation web application with coherent narrative capabilities, deep interaction support and open source technology is proposed. The system is mainly divided into two functions: one is a text generation function that meets the user's storyboard text creation needs and solves the narrative and interaction problems in similar products; the other is an image generation function that meets the user's image preview needs.

[0006] To achieve the above-mentioned purpose, the present invention provides an intelligent storyboard script generation system, comprising a front-end interactive interface layer, a communication layer 1, a back-end business logic layer, a communication layer 2 and a computing layer connected in sequence, wherein the front-end interactive interface layer comprises text editing and image preview functions, the back-end business logic layer comprises a text generation module composed of a large language model and an image generation module composed of a text-to-image generation class model, the text generation module and the image generation module process input content to obtain a computing-intensive task, and the computing layer is set as a cluster composed of a plurality of servers to calculate the computing-intensive task.

[0007] Preferably, the communication layer 1 is set as a Fast api communication layer, and in the communication layer 2, an LLM interface is set for the text generation module, and a Stable-diffusion interface is set for the image generation module.

[0008] Preferably, the text generation module includes a director module, an action script module and a storyboard script module which are connected in sequence. The director module includes two subsidiary modules, namely a role archive module and a scene archive module, which provide role and scene archives to the director module. The action script module converts the output of the director module into an action script, and the storyboard script module converts the output of the action script module into a storyboard script.

[0009] Preferably, the image generation module includes a prompt word generation module, a prompt word completion module and a diffuser workflow module; wherein the prompt word generation module includes an ip-adapter module and a storyboard script conversion module.

[0010] A method for generating an intelligent storyboard script comprises the following steps:

[0011] Step 1: The user provides input content at the front-end interactive interface layer. The input content is the subject, which is input to the back-end business logic layer through the communication layer 1, and the subject is input into the text generation module;

[0012] Step 2: For the text generation module, the module is divided according to function to obtain five agents, namely, the character profile writer, the scene profile writer, the director agent, the action script writer and the storyboard script writer, which specifically correspond to the character profile module, the scene profile module, the director module, the action script module and the storyboard script module. The work of the five agents will generate computing-intensive tasks, which are encapsulated using Docker containerization technology and transmitted to the computing layer for computing through the LLM interface of the communication layer 2;

[0013] A character profile module and a scene profile module are set in the text generation module. The character profile writer and the scene profile writer provide the character appearance and scene to the director module. The theme is input into the director module. After passing through the director agent, the instruction text is obtained. The instruction text is passed through the action script writer in the action script module to obtain the action script. The action script is passed through the storyboard script writer in the storyboard script module to obtain the storyboard script. During the working period, the instruction text, the action script and the storyboard script are respectively fed back to the front-end interactive interface layer for confirmation. The user modifies the instruction text, the action script and the storyboard script, and finally confirms them.

[0014] Step 3: For the image generation module, an agent is set for the prompt word generation module according to the function, and the agent is set as an AI painting prompt word writer. The character appearance output by the character profile writer and the storyboard script finally confirmed by the front-end interactive interface layer are used as the input of the prompt word generation module to obtain preliminary prompt words. The preliminary prompt words are passed through the prompt word completion module to obtain raw image information. The raw image information is passed through the diffuser workflow module to generate images, and the generated images are transmitted to the front-end interactive interface layer for preview. During the working process of the prompt word generation module, the prompt word completion module and the diffuser workflow module, computing-intensive tasks will be generated, which are encapsulated using Docker containerization technology and transmitted to the computing layer for computing through the Stable-diffusion interface in the communication layer 2.

[0015] Step 4: The user gets the generated text and image.

[0016] Preferably, in the text generation module, the specific working process is as follows:

[0017] The director module provides two generation methods. The first is that when the system is initialized, the director module will generate instruction text according to the theme provided by the user; the second is that when continuing to write, the director module will use the previously generated action script as a memory to generate instruction text;

[0018] The action script module provides two generation methods. The first is when the system is initialized. The input of the action script module will come from the instruction text, character appearance and scene provided by the director module. The second is when continuing to write, it also includes the previously existing action script. The action script module converts the instruction text generated by the director module into a series of action scripts. The action script is a series of single sentences that express a single action and visual moment. The action script is saved and sent to the director module as memory;

[0019] The storyboard script module provides two generation methods. The first one is batch generation. During the initial generation, a batch of storyboard scripts corresponding to the action script are generated in the same LLM dialog window according to the batch size. The second one is single generation. After the user edits the action script, the storyboard script is regenerated according to the modified action script.

[0020] Preferably, the specific working process of the image generation module is as follows:

[0021] For the input character appearance, the ip-adapter module is used to perform image reasoning on the content of the character profile module to obtain the character image;

[0022] For the input storyboard script, after the storyboard script conversion module, the storyboard text input is converted into four windows, namely target, expression, camera position, environment and atmosphere. These five fields will be combined into preliminary prompt words with different weights;

[0023] For the prompt word completion module, the raw image information is obtained based on the preliminary prompt words, and the raw image information includes positive prompt words, negative prompt words, number of inference iterations, guidance coefficient, image size, reference image and reference weight;

[0024] For the diffuser workflow module, image inference is performed based on the raw image information and the image of the person to obtain the image.

[0025] Therefore, the present invention adopts the above-mentioned intelligent storyboard script generation system and method, which has the following advantages:

[0026] (1) In the present invention, a sound WEB application framework with development potential and market competitiveness is provided; a system architecture design with front-end and back-end separation is adopted, in which the front-end is mainly responsible for providing a user-friendly interactive interface, and the back-end integrates the core business logic and the computing layer for calculation, and can process tasks with high computing requirements such as text generation and image generation.

[0027] (2) In the present invention, the internal structure is divided into different modules, and each part realizes internal functional integration, which can be abstracted into specific agents in the program. That is, each agent can be generated through a different large language model, so that the large model can be trained and optimized in a targeted manner, thereby improving and specializing the performance.

[0028] (3) In the present invention, all texts generated by the text generation module can be modified through the front-end interactive interface layer, so that the user can control the generation results of each stage, so that the final storyboard script can better meet the needs of the user.

[0029] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a general structural diagram of an intelligent storyboard script generation system of the present invention;

[0031] Figure 2 This is a structural diagram of a text generation module in an intelligent storyboard script generation system of the present invention;

[0032] Figure 3 This is a structural diagram of a picture generation module in an intelligent storyboard script generation system of the present invention;

[0033] Figure 4 It is the Self-BLEU score of the present invention under the first strategy in the evaluation process of the intelligent storyboard script generation system and method of the present invention;

[0034] Figure 5 It is the Self-BLEU score of the control group under the first strategy in the evaluation process of the intelligent storyboard script generation system and method of the present invention;

[0035] Figure 6 The Self-BLEU score of the present invention under the second strategy in the evaluation process of the intelligent storyboard script generation system and method of the present invention;

[0036] Figure 7 It is the Self-BLEU score of the control group under the second strategy in the evaluation process of the intelligent storyboard script generation system and method of the present invention;

[0037] Figure 8 is the average ROUGE score of the indication text and the action script in the intelligent storyboard script generation system and method of the present invention;

[0038] Fig. 9 It is the average ROUGE score of the action script and the storyboard script in the intelligent storyboard script generation system and method of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. The specific model specifications need to be selected and determined according to the actual specifications of the device, and the specific selection calculation method adopts the existing technology in the field, so it will not be described in detail.

[0040] Example

[0041] like Figure 1-Figure 3 As shown, the present invention provides an intelligent storyboard script generation system, comprising a front-end interactive interface layer, a communication layer one, a back-end business logic layer, a communication layer two and a computing layer connected in sequence, the front-end interactive interface layer comprises text editing and image preview functions, the front-end interactive interface layer is developed using a Streamlit framework, is specially designed for rapidly creating data-driven applications, and supports real-time data updates and interactive components, the back-end business logic layer comprises a text generation module composed of a large language model and an image generation module composed of a text-to-image generation class model, the text generation module and the image generation module process input content to obtain a computationally intensive task, and the computing layer is set as a cluster composed of a plurality of servers to compute the computationally intensive task.

[0042] Communication layer 1 is set as Fast api communication layer. In communication layer 2, LLM interface is set for text generation module and Stable-diffusion interface is set for image generation module.

[0043] The text generation module includes a director module, an action script module and a shot script module which are connected in sequence;

[0044] The director module is a zero-shot module, which means that the performance of the module function will depend entirely on the emergence ability of the large language model, which can stimulate the creativity of the director module as a large language model.

[0045] The director module includes two subsidiary modules, namely the role archive module and the scene archive module, which provide the director module with role and scene archives, help the director module to enrich the role and scene settings, and assist the director agent in making plot decisions. At the same time, in order to avoid unclear reference problems, the role archive module and the scene archive module are not designed in the director module according to smooth logic, but are placed in the text generation system as plug-ins as a means of interaction between users and the system;

[0046] The action script module converts the output of the director module into an action script, which very specifically expresses who is, where, what is being done, and which step has been completed. It is a situational description obtained after expanding the plot theme provided by the director, and a text summary of the image objects shot subsequently. Therefore, the action script module is a key transition module between the director and the storyboard script.

[0047] The storyboard script module converts the output of the action script module into a storyboard script;

[0048] The image generation module includes a prompt word generation module, a prompt word completion module and a diffuser workflow module; wherein the prompt word generation module includes an ip-adapter module and a storyboard script conversion module.

[0049] A method for generating an intelligent storyboard script comprises the following steps:

[0050] Step 1: The user provides input content at the front-end interactive interface layer. The input content is the subject, which is input to the back-end business logic layer through the communication layer 1, and the subject is input into the text generation module;

[0051] Step 2: For the text generation module, the module functions are divided into five agents, namely, the role profile writer, the scene profile writer, the director agent, the action script writer and the storyboard script writer, which specifically correspond to the role profile module, the scene profile module, the director module, the action script module and the storyboard script module. The work of the five agents will generate computing-intensive tasks, which are encapsulated using Docker containerization technology and transmitted to the computing layer for computing through the LLM interface of the communication layer 2;

[0052] Set up the role profile module and scene profile module in the text generation module. The role profile writer and scene profile writer provide the role appearance and scene to the director module, and input the theme into the director module. After the director agent, the logical position of the director module in the system is parallel to that of the user. The director module will autonomously process and evaluate the generated content, and then guide the development of the plot. The director module provides two generation methods. The first is that when the system is initialized, the director module will generate instruction text according to the theme provided by the user; the second is that when continuing to write, the director module will use the previously generated action script as a memory to generate instruction text;

[0053] The instruction text is obtained, and the action script is obtained by the action script writer in the action script module. The action script module provides two generation methods. The first is when the system is initialized. The input of the action script module will come from the instruction text, character appearance and scene provided by the director module. The second is when continuing to write, it also includes the previously existing action script. The action script module converts the instruction text generated by the director module into a series of action scripts. The action script is a series of single sentences that express a single action and visual moment. The action script is saved and sent to the director module as memory. The existence of memory can help to continue the plot. Multiple action scripts can be used for the generation of storyboard scripts. The position of the action script in the entire system is very important. It is used to truly undertake the advancement of the plot and is the key to solving the problem of narrative coherence.

[0054] The action script is generated by the storyboard writer in the storyboard script module to obtain the storyboard script. The storyboard script module provides two generation methods. The first is batch generation. During the initial generation, a batch of storyboard scripts corresponding to the action script are generated in the same LLM dialog window according to the batch size. The second is single generation. After the user edits the action script, the storyboard script is regenerated according to the modified action script. Among them, the storyboard script module is a pure expansion module. It will use professional language to convert the action script into a storyboard script in the format required by the user, and expand the condensed one-dimensional information to supplement the information outside the plot such as picture, camera movement, sound, etc.

[0055] During the work, the instruction text, action script and shot script are fed back to the front-end interactive interface layer for confirmation. The user modifies the instruction text, action script and shot script and finally confirms it. The interaction process with the text generation module is as follows: Figure 2 shown.

[0056] Step 3: For the picture generation module, set an agent for the prompt word generation module according to the function, set it as an AI painting prompt word writer, and use the character appearance output by the character profile writer and the storyboard script finally confirmed by the front-end interactive interface layer as the input of the prompt word generation module;

[0057] For the prompt word generation module, for the input character appearance, the ip-adapter module is used to perform image reasoning on the output content of the character profile module to obtain the character image; Ip-Adapter is a text-compatible image prompt adapter, which is mainly used for pre-trained text-to-image diffusion models to achieve the ability to generate images using image prompts. The main functions are: identifying the artistic style and content of the reference image and generating similar works; used in conjunction with text prompts to complete multimodal image generation.

[0058] For the input storyboard script, the storyboard text input is converted into four windows, namely target, expression, camera position, environment and atmosphere. Among them, "target" indicates the main raw image object, "expression" explains the state of the raw image object, "camera position" provides the overall camera position, and "environment and atmosphere" determines the environment and style. The design principle of this prompt word template is to cover more plot information with a shorter prompt word length while ensuring the stability of the entire image. The prompt word template intentionally omits the description of the character image to avoid contaminating the prompt words related to the plot. These five fields will be combined into preliminary prompt words with different weights;

[0059] The preliminary prompt words are passed through the prompt word completion module to obtain raw image information, which includes positive prompt words, negative prompt words, number of inference iterations, guidance coefficient, image size, reference image and reference weight;

[0060] The raw image information and the person image are subjected to image reasoning through the diffuser workflow module to generate images, and the generated images are transmitted to the front-end interactive interface layer for preview; during the working process of the prompt word generation module, the prompt word completion module and the diffuser workflow module, computing-intensive tasks are generated, which are encapsulated using Docker containerization technology and transmitted to the computing layer for calculation through the Stable-diffusion interface in the communication layer 2;

[0061] When selecting the diffuser flow working module, such as Figure 3 As shown in the figure, the basic diffuser workflow can be selected to be built through pod on the basis of runpod-volume, and runpod-volume and serverless container are linked to provide cache and pipeline reading capabilities. After that, the basic diffuser workflow is reused in the runpod container. In addition to the basic inference pipeline, there is also an optional more complex ip-adapter pipeline and optional deepcache activation.

[0062] In the actual process, the most basic diffuser workflow will be built in the cloud and pushed to huggingface for storage. Its model weights will be saved to the volume. The workflow will be used as a pipeline to build more complex pipelines in the Docker container. Finally, the entire inference process will be executed in the Docker container in the runpod serverless service. The serverless endpoint will communicate with the API, accept inference parameters and return inference results.

[0063] Functionally, the diffuser workflow module needs to receive raw image information from the backend business logic layer. The diffuser workflow will perform image inference based on the received parameters, and convert the inferred image into a base64 byte stream and return it, which will be sent to the front-end interactive interface layer for preview.

[0064] Step 4: Based on the storyboard script in step 2 and the image generated in step 3, the user obtains the corresponding generated text and image in the front-end interactive interface layer, and can modify the content and regenerate it.

[0065] The generated content is evaluated as follows:

[0066] Repeatability evaluation is performed. For a WEB application system based on a large language model, whether the repetitive generation of the language model can be effectively avoided is an important evaluation indicator, which will directly affect the narrative ability of the system. In the technical solution of the present invention, multiple language model agents are used to collaboratively complete the text generation task, mainly including three modules: director, action script and storyboard script. The addition of director and action script modules significantly improves the narrative ability of the system and avoids the problem of model repetition itself to a certain extent.

[0067] In the specific evaluation process, the Self-BLEU score is selected to evaluate whether the system has generated repetitive generation. The Self-BLEU score is an indicator used to measure the diversity of generated text. It is based on the calculation method of the BLEU (Bilingual Evaluation Understudy) score. It evaluates diversity by comparing the similarity between each sentence and other sentences in the generated text collection. The BLEU score between each sentence and other sentences can be calculated to obtain the average BLEU score of each sentence. This value reflects the uniqueness of the sentence in the collection. Then, the average BLEU score of all sentences is calculated to obtain the Self-BLEU score. The lower the Self-BLEU score, the better the diversity of the generated text collection, and vice versa.

[0068] In the simulation, we first collected multiple storyboards from the same input theme in the present invention, and each generation task generated 18 related storyboards. Then, this subject simulated the generation content of the model generation and continued storyboard scripts without the director and action script modules through gpt3.5 as the control group. In the control group, 18 related storyboards were also generated in each separate generation task. Next, two strategies were selected to calculate the Self-BLEU scores between the storyboard scripts;

[0069] In the first strategy, the text generation process is simulated, and new storyboards are continuously added to the old text collection to calculate the Self-BLEU score, such as Figure 4-Figure 5 As shown, Figure 4 is the average Self-BLEU score of this system in the generation task, Figure 5 The average Self-BLEU score of the control group without the director and action script module in the generation task can be clearly seen. Figure 4 In the example, the average Self-BLEU score of the storyboard script set tends to a low value of 0.05 as new shots are added. Figure 5 The average Self-BLEU score of the medium-length shot script set increases significantly with the addition of new shots.

[0070] In the second strategy, the present invention measures the similarity between shots by taking three shots as a group, such as Figure 6-Figure 7 As shown, between adjacent shots, the experimental results of the present invention and the control group perform comparable in Self-BLEU scores, but it can also be noted that the control group shows an increasing trend in the Self-BLEU score as the number of shots increases.

[0071] In summary, by setting up a control experiment to calculate the Self-BLEU score, compared with the existing method, the generated content of the present invention is not prone to fall into repetition and looping. The design of the director and action script modules plays a positive role in avoiding repetitive generation and improving the interactive experience of the system.

[0072] Semantic coherence evaluation is performed, and the ROUGE score is used to evaluate the semantic coherence of the system hierarchical structure. The purpose is to evaluate whether the semantics from the instruction text to the action script and then to the storyboard script are consistent and coherent. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a metric for evaluating the similarity between automatically generated text and reference text. ROUGE measures the similarity between the generated text and the reference text by calculating the n-gram coverage. The n-gram coverage refers to the proportion of n-grams (continuous n words) in the generated text that appear in the reference text. Specifically, for each n-gram, it is necessary to check whether it appears in the reference text B. If it appears, the n-gram is considered to be a match.

[0073] Recall and precision are two core indicators of ROUGE, which come from the field of information retrieval. Recall refers to the number of matched n-grams divided by the total number of all n-grams in the reference text B. It measures the proportion of n-grams matched in the generated text to all n-grams in the reference text. Precision refers to the number of matched n-grams divided by the total number of n-grams in the generated text A. It measures the proportion of n-grams matched in the generated text to all n-grams in the generated text.

[0074] To combine recall and precision, ROUGE calculates the F1 score, which is the harmonic mean of recall and precision.

[0075] The present invention calculates the ROUGE scores between the generated content of different modules in the generated content, and counts the ROUGE information of different granularities. ROUGE-1 focuses on the matching of word level, ROUGE-2 focuses on the matching of two-word phrases, and ROUGE-1 focuses on the matching of longer sequences. Figure 8-Figure 9 As shown, from Figure 8It can be observed that the rouge-1 precision of the instruction text and action script modules is higher, which means that the generated text of the action script module grasps some keywords output by the director module, and the two modules have good consistency in entity reference such as nouns;

[0076] from Fig. 9 As can be seen, the average rouge scores between the action script module and the storyboard script module are both high. The precision of the rouge-1 score reaches 0.631, which shows that most of the keywords in the action script are captured by the storyboard script. The precision of the rouge-2 score also exceeds 0.5, indicating that the output sentences of the action script and the storyboard script modules are very similar in terms of two-word phrases.

[0077] In summary, the rouge scores of the content generated by different modules show that the system has relatively coherent semantics between layers and good consistency in word reference.

[0078] Therefore, the present invention adopts an intelligent storyboard script generation system and method, through which a storyboard script generation web application with coherent narrative ability, deep interaction support, and open source technology can be constructed, which can meet the actual creation needs of ordinary users in the self-media era and support further iterative upgrades under the development framework. The system mainly realizes two functions for users, one is the text generation function that meets the user's storyboard text creation needs and solves the narrative and interaction problems in similar products, and the other is the image generation function that meets the user's image preview needs.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. An intelligent storyboard script generation system, characterized by: It includes a front-end interactive interface layer, a communication layer 1, a back-end business logic layer, a communication layer 2 and a computing layer connected in sequence, wherein the front-end interactive interface layer includes text editing and image preview functions, the back-end business logic layer includes a text generation module composed of a large language model and an image generation module composed of a text-to-image generation class model, the text generation module and the image generation module process input content to obtain a computing-intensive task, and the computing layer is set as a cluster composed of a plurality of servers to calculate the computing-intensive task; The communication layer 1 is set as a Fast api communication layer, and in the communication layer 2, an LLM interface is set for the text generation module, and a Stable-diffusion interface is set for the image generation module; The text generation module includes a director module, an action script module and a storyboard script module connected in sequence. The director module includes two subsidiary modules, namely a role archive module and a scene archive module, which provide the director module with role and scene archives. The action script module converts the output of the director module into an action script, and the storyboard script module converts the output of the action script module into a storyboard script. The image generation module includes a prompt word generation module, a prompt word completion module and a diffuser workflow module; wherein the prompt word generation module includes an ip-adapter module and a storyboard script conversion module.

2. An intelligent storyboard script generation method, comprising a storyboard script generation system according to claim 1, characterized in that: The following steps are involved: Step 1: The user provides input content at the front-end interactive interface layer. The input content is the subject, which is input to the back-end business logic layer through the communication layer 1, and the subject is input into the text generation module; Step 2: For the text generation module, the module is divided according to function to obtain five agents, namely, the character profile writer, the scene profile writer, the director agent, the action script writer and the storyboard script writer, which specifically correspond to the character profile module, the scene profile module, the director module, the action script module and the storyboard script module. The work of the five agents will generate computing-intensive tasks, which are encapsulated using Docker containerization technology and transmitted to the computing layer for computing through the LLM interface of the communication layer 2; A character profile module and a scene profile module are set in the text generation module. The character profile writer and the scene profile writer provide the character appearance and scene to the director module. The theme is input into the director module. After passing through the director agent, the instruction text is obtained. The instruction text is passed through the action script writer in the action script module to obtain the action script. The action script is passed through the storyboard script writer in the storyboard script module to obtain the storyboard script. During the working period, the instruction text, the action script and the storyboard script are respectively fed back to the front-end interactive interface layer for confirmation. The user modifies the instruction text, the action script and the storyboard script, and finally confirms them. Step 3: For the image generation module, an agent is set for the prompt word generation module according to the function, and the agent is set as an AI painting prompt word writer. The character appearance output by the character profile writer and the storyboard script finally confirmed by the front-end interactive interface layer are used as the input of the prompt word generation module to obtain preliminary prompt words. The preliminary prompt words are passed through the prompt word completion module to obtain raw image information. The raw image information is passed through the diffuser workflow module to generate images, and the generated images are transmitted to the front-end interactive interface layer for preview. During the working process of the prompt word generation module, the prompt word completion module and the diffuser workflow module, computing-intensive tasks will be generated, which are encapsulated using Docker containerization technology and transmitted to the computing layer for computing through the Stable-diffusion interface in the communication layer 2. Step 4: The user gets the generated text and image.

3. The method for generating an intelligent storyboard script according to claim 2, wherein: In the text generation module, the specific working process is as follows: The director module provides two generation methods. The first is that when the system is initialized, the director module will generate instruction text according to the theme provided by the user; the second is that when continuing to write, the director module will use the previously generated action script as a memory to generate instruction text; The action script module provides two generation methods. The first is when the system is initialized. The input of the action script module will come from the instruction text, character appearance and scene provided by the director module. The second is when continuing to write, it also includes the previously existing action script. The action script module converts the instruction text generated by the director module into a series of action scripts. The action script is a series of single sentences that express a single action and visual moment. The action script is saved and sent to the director module as memory; The storyboard script module provides two generation methods. The first one is batch generation. During the initial generation, a batch of storyboard scripts corresponding to the action script are generated in the same LLM dialog window according to the batch size. The second one is single generation. After the user edits the action script, the storyboard script is regenerated according to the modified action script.

4. The method for generating an intelligent storyboard script according to claim 2, characterized in that: The specific working process of the image generation module is as follows: For the input character appearance, the ip-adapter module is used to perform image reasoning on the content of the character profile module to obtain the character image; For the input storyboard script, after the storyboard script conversion module, the storyboard text input is converted into four windows, namely target, expression, camera position, environment and atmosphere. These five fields will be combined into preliminary prompt words with different weights; For the prompt word completion module, the raw image information is obtained based on the preliminary prompt words, and the raw image information includes positive prompt words, negative prompt words, number of inference iterations, guidance coefficient, image size, reference image and reference weight; For the diffuser workflow module, image inference is performed based on the raw image information and the image of the person to obtain the image.

Images